Can robots be trained to pick up objects they've never seen before? Recent research indicates that the best approach may be to allow robots to experiment on their own, and then share that knowledge with other robots.
How To Train Your Robot
Until now, robots were programmed to recognize specific objects and to relate to those objects in a pre-programmed fashion. This is a good approach for predictable tasks, in planned locations, with previously experienced objects. In the real world, life is what happens while you’re making other plans.
At Google, Sergey Levine and his team took fourteen robotic arms, networked them together, and used (CNN) convolutional neural networks to let these robots learn on their own how to pick up small objects including a cup, tape dispenser, and a toy dolphin.
The robots learned from practice and sharing knowledge with one another, and without human interaction or pre-programming. In fact, the only human contribution was to replace the objects in the trays.
During the study, the robots were positioned in front of trays with randomly placed objects varying in shape, scale, weight and hardness. Each of the 14 robot arms had a similar, but not identical design.
Each robot’s successes and failures were fed to a shared convoluted neural network, a “robo-wiki” of shared knowledge accessible to the 14 robots.
“Our experimental evaluation demonstrates that our method achieves effective real-time control, can successfully grasp novel objects, and corrects mistakes by continuous robotic hand-eye coordination...”
Google’s project builds on the concept that robots can teach other robots how to perform specific tasks—or, more accurately, how one artificial intelligence can help another AI improve, without a human ever intervening.
Here's a link to the Google Research: Learning Hand-Eye Coordination for Robotic Grasping with Deep Learning and Large-Scale Data Collection - Sergey Levine, Peter Pastor, Alex Krizhevsky, and Deirdre Quillen
The robots taught each other to pick up hard and soft objects in different ways, without human guidance.
For an object that was perceived as rigid the robots would just grasp the outer edges and squeeze to hold it tightly. But for a soft object, like a sponge or the dolphin, the robots "learned" that it was easier to place one gripping "finger" in the middle and one around the edge, and then just squeeze.
The good news -- the Robots still needed humans to place objects in the trays.
A Brown University project called “RoboBrain,” plans to store all machine-learned behaviors into a knowledge database which other bots can tap into. The project has been nicknamed “the world’s first knowledge engine for robots.” Some people call in "SkyNet."
Robo Brain is a large-scale computational system that learns from publicly available Internet resources, computer simulations, and real-life robot trials.
It accumulates everything robotics into a comprehensive and interconnected knowledge base. Applications include prototyping for robotics research, household robots, and self-driving cars.
The Claw Is Our Master: He chooses who will go and who will stay.
(c) David J. Katz, 2016 - New York City